PAC-FNO: Parallel-Structured All-Component Fourier Neural Operators for Recognizing Low-Quality Images

Authors: Jinsung Jeon, Hyundong Jin, Jonghyun Choi, Sanghyun Hong, Dongeun Lee, Kookjin Lee, Noseong Park

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensively evaluating methods with seven image recognition benchmarks, we show that the proposed PAC-FNO improves the performance of existing baseline models on images with various resolutions by up to 77.1% and various types of natural variations in the images at inference.
Researcher Affiliation Academia 1Yonsei University, 2Seoul National University, 3Oregon State University, 4Texas A&M University-Commerce, 5Arizona State University, 6KAIST
Pseudocode Yes We describe a two-stage training algorithm in Algorithm. 1. We first stage for training with target resolution in the first while loop, followed by the second while loop for training with various resolutions. Algorithm 1: 2-stage Training of PAC-FNO
Open Source Code Yes For reproducibility, we attached the source codes in our supplementary materials.
Open Datasets Yes Datasets. We use seven image recognition benchmark datasets to evaluate PAC-FNO. For lowresolution tasks, six image recognition benchmark datasets are used to evaluate PAC-FNO: Image Net-1k (Russakovsky et al., 2015), Stanford Cars (Krause et al., 2013), Oxford-IIIT Pets (Parkhi et al., 2012), Flowers (Nilsback & Zisserman, 2008), FGVC Aircraft (Maji et al., 2013), and Food-101 (Bossard et al., 2014). For input variation task, we use Image Net-C/P (Hendrycks & Dietterich, 2019) which is an image dataset containing 19 common corruptions and perturbations.
Dataset Splits Yes For training we use {224} or {299} as target resolution depending on the backbone model along with {32, 64, 128} low resolution. ... The number of classes, train images, and test images are organized in Table 7. Image Net-1k 1000 1,281,167 50,000
Hardware Specification Yes We run our experiments on a machine equipped with Intel i9 CPUs and Nvidia RTX A5000/A6000 GPUs.
Software Dependencies Yes We implement PAC-FNO using Python v3.8 and Py Torch v1.12.
Experiment Setup Yes Detailed setup is found in Appendix F.1. ... In Tables 9 and 10, we list all the key hyperparameters in our experiments for each dataset.